基于改进RT-DETR的航拍图像小目标检测算法OA

To address the problems of low detection accuracy,high model parameters,and insufficient feature extraction in complex backgrounds for small object detection in aerial images,a quick-enhanced real-time detection transformer(QRT-DETR)model for small object in aerial images was proposed.Built upon real-time detection transformer(RT-DETR),the model introduced a lightweight ContextFocus Net in the backbone,whose core ELGC-Mixer module employed a dual-path structure with a gating fusion mechanism to extract and fuse local details and global context information separately.In the feature fusion network,a high-frequency enhanced cross-scale fusion module was designed,utilizing a high-resolution P2 layer and parameter-free SimAM attention to construct a dedicated high-frequency information pathway.A joint loss function combining shape-aware Inner-ShapeIoU with normalized Wasserstein distance was introduced to optimize bounding box regression.Ablation experiments and performance comparisons with other models were conducted on the VisDrone2019 dataset.The results show that QRT-DETR achieves 52.0％mAP@50 and 32.6％mAP@50-95,improving by 5.1 percentage pointsand 4.0 percentage points over the original RT-DETR,respectively,while reducing parameters by 41.8％.The proposed model effectively enhances the detection accuracy for small objects in complex scenes and achieves lightweight design,providing a feasible technical solution for real-time small object detection in UAV aerial imagery.